Hy4 preview is positioned as a next-generation productivity model in the Hy family, with Tencent emphasizing significantly enhanced agent capabilities and complex task execution. The architecture is a Mixture-of-Experts design on a 78-layer backbone, totaling 770B parameters with 49B activated per token, using 256 routed experts plus a shared expert per layer and selecting 8 routed experts per forward pass. Open weights are published, which lets teams self-host in addition to accessing the model through hosted endpoints, an unusual combination at this parameter scale and a meaningful signal for organizations that need on-premise deployment for privacy or compliance reasons.
In benchmark rankings aggregated by LLM-Stats on release, Hy4 preview lands in the top decile for Tool Calling (7 of 180), Reasoning (8 of 346), Coding (10 of 253), and Math (12 of 316), with a WideSearch score of 0.84 that ranks first among tracked models and a Terminal-Bench 2.1 score of 0.85 placing eighth. Specific reported scores include GDPval-AA at 1678/3000 (rank 4) and GPQA Diamond at 0.92, though some figures come from provider-supplied scorecards. Target workloads highlighted at launch include long-horizon coding, tool-using agents, office analysis, game development, and scientific tasks, with software engineering framed as a central use case, making the model a strong fit for teams building autonomous coding or research pipelines that need a large open-weights model with strong reasoning and tool use.
